IC-1601GPT-4's self-generated feedback is significantly less effective than human programmer feedback for code repair, with the gap widening on harder problems

Theo X. Olausson, Jeevana Priya Inala, Chenglong Wang, Jianfeng Gao, Armando Solar-Lezama

SourceIs Self-Repair a Silver Bullet for Code Generation?

In a small-scale study with 16 human participants providing feedback on 40 failing GPT-4 programs, replacing GPT-4's own feedback with human feedback increases the fraction of repaired programs passing all unit tests from 33.3% to 52.6% (1.58x). The relative improvement is larger on harder problems: introductory 42.64% to 62.21%, interview 19.33% to 45.67%, competition 3.67% to 14.67%. Qualitative analysis shows GPT-4's feedback is more often obviously inaccurate (32/80 vs 7/80), less likely to suggest high-level changes, and never expresses uncertainty (0/80 vs 7/80).

Evidence
correlational
Key metric
Overall repair success: GPT-4 feedback 33.30% vs human feedback 52.60% (1.58x). By difficulty: introductory 42.64% vs 62.21%, interview 19.33% vs 45.67%, competition 3.67% vs 14.67%. GPT-4 obviously inaccurate feedback: 32/80 vs 7/80 human.
Caveat
Small sample: 40 programs, 16 participants (15 graduate students, 1 professional ML engineer). Participants were not timed, so the cost of human feedback is not accounted for. The study was approved by IRB and carried out via online survey.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Concepts
Failure mode
Datasets
APPS [eval]
Related findings
IC-1599, IC-1600
Extraction
automatic-extraction